NLP Development in Zurich, Switzerland

Insurance / Group Operations

A Zurich-based insurance group (CHF 5.8 billion GWP, operating across Switzerland, Germany, Austria, and Liechtenstein, 12,000 employees) processed approximately 2.4 million incoming documents annually across its operations — claims submissions, policy applications, broker correspondence, customer enquiries, regulatory communications, and internal documents. Documents arrived in: German (62%), French (18%), Italian (8%), and English (12%) — via email (55%), postal mail/scan (30%), portal upload (12%), and fax (3%). The document processing challenge: incoming documents were manually sorted by a 35-person mailroom and document management team. Each document was: opened/scanned (for postal mail), classified by type (claim, application, enquiry, correspondence, regulatory), classified by business unit (motor, property, liability, health, life), classified by urgency (routine, urgent, regulatory deadline), and routed to the appropriate department and handler. Manual classification was: slow (average 48 hours from receipt to handler — with backlogs extending to 5+ days during peak periods), error-prone (12% misclassification rate — documents routed to the wrong department or handler, requiring re-routing and adding 2-3 days to processing time), and expensive (35 FTE at an average cost of CHF 85,000 per employee — CHF 2.975 million annually in document management staffing). The misclassification problem was particularly acute for: multilingual documents (a French-language claim from a Romandie policyholder was occasionally routed to the German-language claims team, causing delays), ambiguous documents (a letter that was both a complaint and a claim notification — requiring routing to both customer relations and claims), and regulatory correspondence (FINMA letters requiring immediate senior management attention were occasionally classified as routine correspondence and delayed).

We built a multilingual document intelligence system over 14 weeks. Document ingestion pipeline (weeks 1-4): the system processed all incoming document channels: email: automated extraction of email body text and attachments (PDF, Word, image). The system processed both the email content and attached documents — understanding that an email saying "Please find attached our claim for the incident on 15 March" contained both the email metadata and the claim document attachment. Scanned mail: integration with the existing scanning workflow — scanned documents entered the pipeline as PDF images. OCR processing using a model fine-tuned on Swiss insurance documents: recognising insurance-specific terminology, Swiss address formats, policy number patterns, and handwritten annotations. Portal uploads: documents uploaded by brokers and customers through the group's online portal — already digital, with metadata from the upload form (document type self-classification, policy number, contact details). Text extraction: all documents converted to structured text with: document content (full text), document metadata (date, sender, reference numbers), document structure (sections, paragraphs, tables), and embedded data (policy numbers, claim numbers, dates, amounts — extracted using regex patterns and NER models trained on Swiss insurance document formats). Classification system (weeks 3-8): multi-label classification: each document was classified across four dimensions simultaneously. Document type (18 categories): claim notification, claim update, claim documentation (repair quote, medical report, police report), policy application, policy amendment request, policy cancellation, coverage enquiry, premium enquiry, complaint, regulatory correspondence, broker communication, legal correspondence, payment query, address change, power of attorney, court correspondence, and general correspondence. Business unit (7 categories): motor, property, liability, health, life, corporate/commercial, and group-level/regulatory. Urgency (4 levels): critical (regulatory deadline, court deadline, complaint escalation), urgent (new claim requiring immediate acknowledgement, time-sensitive customer request), standard (routine processing within SLA), and low (informational, no action required). Language: German, French, Italian, English — with confidence score (some documents contained multiple languages or ambiguous language signals). The classification models were: trained on 180,000 historically classified documents (3 years of classified incoming mail — de-identified where containing personal data), implemented as an ensemble: a multilingual transformer model (XLM-RoBERTa fine-tuned on the insurance corpus) for content-based classification, combined with a rule-based classifier for pattern-matching (policy number formats indicating business unit, sender email domains indicating regulatory or legal correspondence), and confidence scoring (each classification included a confidence score — high-confidence classifications proceeded automatically; low-confidence classifications were routed for human review). Entity extraction (weeks 5-10): from classified documents, the system extracted: policy numbers (matching the group's policy number formats across all business units), claim numbers (for documents relating to existing claims — linking to the claims management system), dates (incident dates, document dates, deadline dates — parsed from German, French, Italian, and English date formats), monetary amounts (CHF, EUR — with context: repair costs, claim reserves, premium amounts, settlement offers), person names (claimants, policyholders, third parties, legal representatives — using NER models trained on Swiss naming conventions), and addresses (Swiss address format — street, PLZ, city, canton — validated against the Swiss address database). Routing and workflow integration (weeks 8-14): automated routing: classified and extracted documents were automatically routed to: the correct business unit queue (motor claims team, property underwriting, health claims, etc.), the correct handler within the team (based on: portfolio assignment for policy-related documents, claim assignment for claims-related documents, or round-robin for new items), with all extracted entities pre-populated in the claims or policy management system (the handler received the document with policy number, claim number, dates, and amounts already entered — rather than reading the document and manually entering data). Human review queue: documents with low classification confidence (below 0.85 threshold) were routed to a human review queue with: the AI classification recommendation (and confidence score), the extracted entities (for validation), and the source document (for the reviewer to verify). The reviewer confirmed or corrected the classification — their decision fed back into the training data for continuous model improvement. Integration: the system integrated with the group's document management system (OpenText) for archival, the claims management system (Guidewire ClaimCenter) for claims-related documents, and the policy administration system for policy-related documents.

The document intelligence system deployed in phases — email channel first (55% of volume), then scanned mail (30%), then portal uploads (12%). Classification accuracy: overall classification accuracy: 94.2% (across all four classification dimensions — document type, business unit, urgency, and language). By dimension: document type: 92.8% (18 categories — some categories with subtle distinctions between types, e.g., claim notification vs. claim update), business unit: 96.1% (7 categories — policy and claim numbers provided strong signals), urgency: 91.4% (4 levels — urgency was the most subjective classification and benefited most from human review), and language: 99.1% (straightforward for clean text; occasional errors on very short documents or documents mixing languages). The 5.8% requiring human review were efficiently processed: the reviewer validated or corrected the AI classification in an average of 45 seconds per document (compared to 3.5 minutes for full manual classification). Processing speed: document-to-handler time reduced from 48 hours average to 4.2 hours (91% reduction). During peak periods (previously 5+ day backlogs): the AI system processed documents at consistent speed regardless of volume — eliminating backlogs entirely. Regulatory correspondence (FINMA, court deadlines): now flagged and escalated within 30 minutes of receipt — previously, critical regulatory letters occasionally waited 2-3 days in the manual classification queue. Misclassification reduction: misclassification rate dropped from 12% to 2.1% — an 83% reduction. The remaining 2.1% were predominantly borderline cases (documents legitimately spanning multiple categories) rather than clear errors. Re-routing delays (previously adding 2-3 days per misclassified document) decreased proportionally. Entity extraction accuracy: policy number extraction: 97.3% (high — structured format easily recognised). Monetary amounts: 95.8%. Dates: 94.1% (lower due to ambiguous date formats and contextual date references). Person names: 91.2% (Swiss naming conventions with particles and double surnames created some challenges). Staffing impact: the 35-person document management team was restructured: 12 positions redeployed to higher-value roles (claims investigation support, quality assurance, customer communication — roles that required human judgment rather than classification), 8 positions assigned to AI review queue management and continuous improvement (reviewing low-confidence classifications, correcting errors, and curating training data for model improvement), and 15 positions reduced through natural attrition over 18 months (retirements and voluntary departures — no forced redundancies). Financial impact: staffing efficiency: CHF 1.275 million annually (15 positions at average CHF 85,000). Processing speed improvement: estimated CHF 800,000 in accelerated claims processing and customer response (faster document routing led to faster claims settlement and customer service). Misclassification reduction: estimated CHF 340,000 in eliminated re-routing delays, customer complaints from delayed processing, and regulatory risk from delayed regulatory correspondence. Total annual benefit: CHF 2.415 million. Implementation cost: CHF 420,000. First-year ROI: 4.8x. Ongoing cost: CHF 7,500/month (infrastructure, model updates, continuous training).